Examples steer a model more reliably than descriptions

Start here

A reader who keeps rewriting a description of what they want, instead of showing an example of it, is missing a shortcut that a foundational paper and two current vendor guides all point to. Rather than a single vendor's claim, three separate documents agree: showing a model two or three worked examples often steers its output further than a longer description of the task does.

What the documents say

The term comes from OpenAI's 2020 paper Language Models are Few-Shot Learners, submitted 28 May 2020, which describes few-shot use as giving a 175-billion-parameter model, GPT-3, task demonstrations purely through text, with no gradient updates or fine-tuning at all: the model is not retrained, only shown examples inside the prompt. The paper reports that this in-context approach reached performance sometimes competitive with prior fine-tuned, state-of-the-art approaches on a range of language benchmarks, a research claim about that model and those benchmarks, not a promise for every task. Anthropic's current prompting reference calls examples one of the most reliable ways to steer Claude's output format, tone and structure, and recommends three to five of them, wrapped in explicit tags. OpenAI's current guide lists few-shot learning as a named technique for guiding a task without fine-tuning, the same distinction the 2020 paper drew.

Check this

The check that catches the failure mode is in the vendor guidance itself, not a separate study: Anthropic's page says examples should be diverse so the model does not pick up an unintended pattern. A reader can test this by writing three examples that happen to share an accidental trait, all short, all in the same tense, all ending the same way, and checking whether the model's new output copies that trait instead of the task.

What holds and what fails

Examples hold up well for steering format and tone, matching a template, a length, a structure, because that is what a demonstration shows directly. They fail, or mislead, when the examples are too similar to each other, since the model has no way to tell an incidental similarity from an intended rule; both current guides warn about exactly this in their own words, not as an aside. Whether examples ever substitute for a capability a model lacks is a separate question the 2020 paper does not settle for every task, and no source here claims they do.

• Replace one paragraph of task description with two or three worked examples and compare the outputs.
• Check every example for an accidental shared trait before assuming the model will only learn the intended one.
• Wrap examples in a clear boundary, a tag or a heading, so the model can tell them apart from instructions.

The lesson holds across three separate documents, a research paper and two competitors' guides that do not cite each other, which is closer to convergent evidence than one source repeated three times.